Ghost Detections in People‑Counting Sensors: Causes & Fixes

How to test and calibrate for ghost reduction

Preparation

• Log sensor data and a synchronized ground‑truth record (manual counter or short video) for offline review.
• Mark test paths and positions on the floor for repeatability.

Recommended tests

• Single‑person walk: one person walks standard paths at normal speed, repeated both directions.
• Stationary person: person stands at multiple fixed points for dwell tests.
• Crossing pairs: two people cross paths to test occlusion behavior.
• Reflective material check: introduce temporary reflective panels and note false detections.
• Empty room baseline: run sensors for 30–60 minutes with no people to measure inherent noise and false positive rate.

Collect at least 10 repeats per test and record timestamps for each event.

Metrics to measure

• False positive rate: ghosts per hour or per minute.
• False negative rate: missed real people.
• Precision and recall for counts over time windows.
• Track stability: average duration and split/merge events per crossing.

Practical mitigation strategies

Mitigations fall into hardware, placement, and software categories. Use a combination for best results.

Hardware and placement

• Avoid direct alignment with reflective surfaces. Reorient sensors or add non‑reflective shields.
• Raise or lower mounting height to change field geometry and reduce ground reflections.
• Use physical baffling or matte paint in problem zones to reduce multipath returns.
• For radar/depth systems, consider absorptive materials or curtains to dampen strong reflections.

Sensor fusion and redundancy

• Combine complementary sensors (thermal + depth, radar + thermal) and use majority or weighted voting to suppress single‑sensor ghosts.
• Use ingress/egress logic across neighboring sensors: only count an entry when a clear path of sequential activations matches an expected trajectory.

Algorithmic fixes

• Smoothing and temporal filters: require minimum dwell time or persistent trajectory before incrementing counts.
• Gating and validation: apply spatial and velocity gates that reject detections outside realistic human motion parameters.
• Improved clustering: tune clustering distance and temporal window to avoid splitting single targets into multiple tracks.
• Background learning: use adaptive background subtraction to ignore static heat sources or persistent noise patterns.
• Track association: implement robust track management to merge brief split tracks and suppress transient single‑frame detections.

Operational tuning

• Lower sensitivity during known high‑noise periods (e.g., HVAC cycles) or use adaptive thresholds tied to measured noise floor.
• Schedule periodic re‑calibration after environmental changes (new furniture, renovations, or seasonal reflectivity changes).

Validation and ongoing monitoring

After adjustments, validate with the same structured tests and compare results to baseline.

Best practices

• Maintain a regular test cadence (monthly or after changes) and log results.
• Use automated alerts for sudden spikes in ghost rate to trigger inspection.
• Record configuration versions alongside test runs to link changes to outcomes.

Recommended acceptance criteria (example)

• False positives below a target threshold (e.g., <0.5 ghosts per 100 person‑minutes) for the given environment.
• Precision and recall above desired levels for your use case (e.g., >95% for safety applications, >85% for space analytics).

How heat-based (thermal) sensing reduces certain ghost types

Thermal, anonymous heat‑based sensors detect human presence via emitted or reflected heat patterns rather than detailed imagery or RF reflections. That gives several advantages for ghost mitigation and privacy.

Advantages

• Lower multipath sensitivity: thermal energy is less prone to long‑distance multipath reflections that plague radar and some depth systems, reducing phantom echoes.
• Simpler foreground separation: a heat source contrasted against room background is easier to model and track, which helps reduce short transient detections.
• Privacy‑first data: thermal sensors typically produce abstracted heat maps or occupancy counts rather than identifiable images, enabling analytics without personal data.

Limitations

• Thermal sensors can still see false detections from warm equipment, sunlight, or HVAC outlets if not accounted for in background models.
• Thermal resolution and field of view determine sensitivity to small or partially occluded people.

When deploying thermal systems, apply the same calibration and validation steps described earlier. Heat‑based sensing often reduces some ghost sources but is most effective when combined with robust algorithms and thoughtful placement.

Quick checklist for field teams

• Survey the site for reflective surfaces and HVAC outlets before sensor placement.
• Run baseline empty‑room logging to measure noise floor.
• Perform structured tests (single walk, stationary, crossings) and record ground truth.
• Tune clustering, dwell thresholds, and gating conservatively; prefer stability over hyper‑sensitivity.
• Consider sensor fusion for high‑accuracy applications.
• Schedule regular re‑calibration and automated monitoring for ghost spikes.

FAQ

What causes ghost targets in people‑counting systems?

Ghosts come from multipath reflections, reflective surfaces, sensor noise, poor algorithm tuning, occlusion, or external interference.

How can I reduce false positives in radar/depth sensors?

Reorient sensors, add absorptive materials, tune sensitivity and clustering, and use temporal gating and track association to reject brief noise.

Are thermal sensors immune to ghost detections?

No sensor is fully immune, but thermal sensing reduces certain ghost types because thermal energy is less prone to problematic reflections and often yields cleaner foreground separation.

How do you validate people‑counting accuracy in a lab?

Use repeatable path tests, synchronized ground‑truth logging, and compute precision/recall and false positive/negative rates across test scenarios.

What are the fastest wins to reduce ghosts in an operational deployment?

Reposition sensors away from reflective surfaces, enable dwell/time gating, and implement basic track smoothing to ignore transient single‑frame detections.

Closing note

Ghost reductions come from iterative testing: observe the problem, test controlled scenarios, apply hardware and algorithmic fixes, and validate results. Combining careful placement, conservative algorithm settings, and sensor choices geared to the environment — such as heat‑based, privacy‑first sensors in sensitive spaces — will yield the most reliable people‑counting performance over time.